Why not call just PySequence_Tuple() (which is a no-op for tuples)? This would simplify the setup and the cleanup code.

It may be also worth to use PySequence_Fast() and PySequence_Fast_GET_ITEM() in the loop. This will save you allocating a new tuple. For general iterable PySequence_Tuple() creates a list and a tuple, but PySequence_Fast() creates just a list.

But you will need to call PySequence_Fast_GET_SIZE() in a loop because list's size can be changed when you convert items to the C double.

I didn't want any external function calls on the fast path for the common case.

I did try it out. The patch was shorter but it caused a 15% performance degradation:

# baseline
$ pytime -r11 -s 'from math import dist' 'dist((10.5, 12.5), (13.5, 8.5))'
5000000 loops, best of 11: 57.3 nsec per loop

# patched to always call PySequence_Tuple() and always Py_DECREF on exit
$ pytime -r11 -s 'from math import dist' -s 'p, q = (10.5, 12.5), (13.5, 8.5)' 'dist(p, q)'
5000000 loops, best of 11: 64.7 nsec per loop